Backend Engineer · Davis, CA · Available Jun 2026
Hi, I'm
I build backend systems that hold up under LOAD.
Distributed caches, event-driven pipelines and LLM infrastructure serving millions of requests in production.
3 yrs @ Target Corp · Agent observability @ Pilotcrew AI · MS ECE @ UC Davis
The stack I build on
OTLP telemetry ingest, AGENT→LLM→TOOL span waterfalls and cost rollups for LLM agents.
Dependency transplant engine that rewrites vulnerable call sites and verifies the result.
Stop panning and a settle-detector writes an AI area summary with re-geocoded places.
Seven systems I designed, shipped and measured
2025–2026 · Pilotcrew AI · Solo-owned service
Production observability for LLM agents, built from my own PRD. OTLP/OpenInference telemetry ingest, AGENT→LLM→TOOL span waterfalls, token-cost derivation, and p95/error/cost time-series rollups at four granularities, with non-blocking emission that can never slow an agent run.
2026 · AI × Supply-chain security · Autonomous pipeline
A dependency transplant engine for vulnerable packages. Scans a GitHub repo into a dependency + call-site impact graph (catching aliased imports grep misses), fires a mock CVE incident, then autonomously rewrites axios call sites to a behavior-preserving fetch wrapper, verified by sandboxed build/test checks, behavioral comparison, and a multi-judge LLM review with live SSE pipeline streaming. Ports-and-adapters architecture runs fully offline on fake adapters.
2026 · Desktop app · Local-first AI
A focus browser that holds you to one task. You set a single task when the browser opens, it is pinned for the whole session, and one relevance judgment is applied across three surfaces: recommendation feeds on any site, cross-domain navigation, and page loads. A bundled MiniLM model decides most items on device; Groq is consulted exactly twice, to expand the task and to break genuine ties. Every hidden item keeps a one-click "show anyway", and the engine fails open on outage but closed on ambiguity, so it never bricks the web.
2026 · Open source · Agent interoperability
Moves useful working context from Agent A to Agent B. A local-first CLI that turns an agent transcript into a portable, inspectable .ctx file, then compiles it into context another agent can actually continue from, under a token budget. Reads Claude Code transcripts and generic JSON Lines; the conversation never leaves your machine unless you opt into model-based extraction. The offline example hands a coding session to a different model and checks it recovers the exact facts that usually vanish: encodings, row counts, forbidden paths, and the approach that was already rejected.
2026 · Open source · Developer tooling
Two published plugins. multiagent-build: an /orchestrate command coordinating four subagents (architect, coder, tester, tracker) through TDD in strict dependency order. discuss-mode: hook-based guardrail that blocks all file-mutating tools for a turn via PreToolUse deny, scoped by hashed session markers.
2026 · Full stack · AI backend
The map is the input: stop panning and a debounced settle-detector triggers an AI-written area summary with 5 re-geocoded places. Provider-agnostic LLM layer (Gemini/Groq/Ollama) behind one contract, versioned prompts with JSON-schema-constrained output, and five geospatially-bucketed caches (~5 km grid, 30-day TTL) slashing redundant AI calls. Fully offline test suite.
2022–2025 · Target Corporation · Production
The backend behind Target's store lookup at national scale. Pre-computed 3,000+ store distances into RocksDB for sub-millisecond reads, with a Kafka cache-invalidation pipeline hot-swapping MongoDB snapshots. P95 latency down 70%, CPU down 60%, zero downtime on data refreshes.
The through-line in everything I ship: backend systems where latency, throughput, and cost are measured, then improved. Every one of these ran in production.
Pre-computed lookups in RocksDB replacing on-the-fly computation, for sub-millisecond reads across 3,000+ entities.
P95 −70% · CPU −60%Kafka-driven cache invalidation with MongoDB → RocksDB hot-swaps; real-time log processing across enterprise systems.
Zero-downtime deploysA single GraphQL layer standardizing data retrieval across Target: schema design, resolvers, caching, versioning.
30+ internal servicesAgent observability (traces, spans, cost), automated evaluation pipelines, and provider-agnostic LLM backends.
OTLP · Evals · Multi-providerAutomated S3 pipelines with daily refresh cycles replacing manual workflows and third-party data providers.
Infra cost −95%Proximity search, ZIP-to-store intelligence, and multi-stop route planning. The domain where I learned to scale backends.
All U.S. marketsRemote · Davis, CA · LLM Infra
Designed and built an LLM-agent observability platform from PRD to production. OTLP/OpenInference telemetry normalized into AGENT → LLM → TOOL span waterfalls.
Bengaluru, India · 3 years · Production
Re-architected the store-proximity API around pre-computed RocksDB lookups and a Kafka cache-invalidation pipeline, and replaced Microsoft MapPoint with an in-house platform.
Bengaluru, India · Solo tech lead
I was the entire engineering team, system design through deployment. Built an artist/venue marketplace from scratch with real-time chat, Stripe payments and full auth.
Bengaluru, India · HAL
Developed signal conversion algorithms for the Health and Usage Monitoring System (HUMS) and optimized the aircraft diagnostic processing pipeline.
Highlighted chips are what I reach for first
7-qubit ALU on IBM quantum hardware, 16-qubit simulation with Qiskit.
94% accuracy traffic classification, 20% reaction time improvement.
Distributed ledger with real-time data immutability for educational records.
University of California, Davis
Dr. Ambedkar Institute of Technology
Open to backend & SDE roles, new grad June 2026, will relocate. If you're hiring, collaborating, or scaling a backend, the payload box is right below.
Hiring, collaborating, or scaling a backend? Payload below. P95 response time: same day.
Status